Re: spinlock contention of files->file_lock

[Eric Dumazet]

From: Eric Dumazet <edumazet@xxxxxxxxxx>

On Mon, 2013-09-30 at 18:44 -0700, Linus Torvalds wrote:

> Now, that only gets rid of fd_install(), but I suspect you could do
> something similar for put_unused_fd() (that one does need cmpxchg for
> the "next_fd" thing, though). We'd have to replace the non-atomic
> bitops on open_fds[] with atomic ones, just to make sure adjacent bit
> clearings don't screw up concurrent adjacent bit values, but that
> looks fairly straightforward too.

While looking at this (exciting) stuff, I found following bug.

Maybe I am missing something obvious ?

Thanks

[PATCH] fs: fix a race in do_close_on_exec()

commit 6a6d27de ("take close-on-exec logics to fs/file.c, clean it up a
bit") added a possible race, as another thread could resize file table
once we released files->file_lock.

We must reload fdt after getting the lock.

Signed-off-by: Eric Dumazet <edumazet@xxxxxxxxxx>
---
 fs/file.c |    5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/fs/file.c b/fs/file.c
index 4a78f98..b614f13 100644
--- a/fs/file.c
+++ b/fs/file.c
@@ -616,10 +616,10 @@ void do_close_on_exec(struct files_struct *files)
 
 	/* exec unshares first */
 	spin_lock(&files->file_lock);
+	fdt = files_fdtable(files);
 	for (i = 0; ; i++) {
 		unsigned long set;
 		unsigned fd = i * BITS_PER_LONG;
-		fdt = files_fdtable(files);
 		if (fd >= fdt->max_fds)
 			break;
 		set = fdt->close_on_exec[i];
@@ -639,6 +639,9 @@ void do_close_on_exec(struct files_struct *files)
 			filp_close(file, files);
 			cond_resched();
 			spin_lock(&files->file_lock);
+
+			/* We released files->file_lock, we must reload fdt */
+			fdt = files_fdtable(files);
 		}
 
 	}


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/